Neuroscientists sort neurons into types such as fast-spiking parvalbumin cells or somatostatin cells. Their classifiers keep failing in the same places: cells that sit between two types, or that vary smoothly within one type. Geneticists face a different shortfall. The common DNA variants found by genome-wide association studies (GWAS, scans that test millions of single-letter DNA differences against a disease one at a time) explain only part of what twin studies say is inherited. This proposal asks whether these two shortfalls are related. One possibility is that the genes which make a neuron hard to classify are the same genes where much of the unexplained inherited risk for psychiatric disorders sits. These are the genes that slide a cell continuously along a gradient of electrical behaviour, as opposed to the marker genes that decide which family it belongs to.
The test uses existing data. We would use paired recordings of gene expression and electrical behaviour from single mouse neurons (Patch-seq) to split genes into two groups. "Identity genes" separate cell families. "Gradient genes" predict where a cell falls along a continuous electrical spectrum within its family. We would then map both groups to human genes and measure how much heritability of schizophrenia, bipolar disorder and autism falls on each, holding fixed the usual confounds such as gene length, expression level and intolerance to mutation. If gradient genes carry more heritability per gene than identity genes, classification ambiguity stops being a nuisance for cell typing. It becomes a map of where small polygenic effects act.
Core Claim
Common-variant risk for neuropsychiatric disorders should act mostly by nudging a neuron's position along its within-family electrical gradient, not by changing which family it belongs to. Large changes to cell identity are developmentally drastic and should be selected against. Small shifts in excitability are the kind of effect expected from thousands of variants that each fall below the genome-wide significance threshold. The prediction is concrete: per-gene heritability enrichment should be larger for gradient genes than for identity genes, after matching the two groups on mutational constraint, expression, specificity and length. The reverse result, with identity genes winning, would also be informative. It would say that cell-type composition and specification, not tuning of physiology, is the main route of risk.
Where this sits in prior work
Three papers motivate the split. Phenotypic variation of transcriptomic cell types in mouse motor cortex profiled 1,329 neurons with Patch-seq. It found discrete morpho-electric clusters at the family level but continuous variation within families, including an Sst Pvalb Calb2 "bridge" type. It also showed that a sparse linear model on 25 genes, mostly ion channels, predicts electrophysiology with cross-validated $R^2 = 0.38$. Sparse Bottleneck Networks for Exploratory Analysis and Visualization of Neural Patch-seq Data reached $R^2 \approx 0.40$ with a nonlinear encoder that picks out channel genes such as Kcnc2. Classifying Neuronal Cell Types Based on Shared Electrophysiological Information from Humans and Mice reported that its locally sparse classifier confuses Sst+ with Htr3a+|Vip– cells because of real overlap in membrane properties. None of these three ever asks whether the genes behind this continuous variation matter for human disease. They treat the gradient as an object for description or a source of error. We treat it as a gene annotation.
Single-neuron models linking electrophysiology, morphology and transcriptomics across cortical cell types supplies part of the method and a validation target. Its fitted conductance vectors classify subclasses at 69% accuracy, and they agree with ion-channel transcript levels (Kv3.1 across interneurons, $I_h$ with Hcn expression). We use these conductance–gene correspondences to check that our gradient genes are biophysically meaningful, not artefacts of the regression.
On the genetics side, Cross-ancestry, cell-type-informed atlas of gene, isoform, and splicing regulation in the developing human brain is the closest bridge. It prioritizes mechanisms for about 60% of loci across five neuropsychiatric disorders and shows that isoform regulation mediates more heritability than total gene expression. It also constrains our design. Its phenotypes stop at RNA, so it cannot say whether a prioritized gene acts on excitability. Its isoform result also warns that a gene-level annotation may underestimate a channel gene's role. A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala shows that gene modules defined by an intermediate phenotype can carry polygenic risk. Its intermediate phenotype, however, is MRI grey-matter density, which mixes every cell type. The pathway-aggregation papers, Pathway analysis of genome-wide data improves warfarin dose prediction and A pathway-based association analysis model using common and rare variants, establish the general logic that weak signals pooled over biologically defined gene groups recover variance missed by single-variant tests. Our gene groups are defined by measured single-cell physiology, not by curated pathway membership.
Research Program
The first stage builds the two annotations from mouse Patch-seq. For each neuron $i$, let $y_i$ be its vector of standardized electrophysiological features, $x_i$ its log expression vector, and $f_i$ its family label. We fit
$$y_i = B\, f_i + W^\top (x_i - \bar{x}_{f_i}) + \epsilon_i ,$$
where $B$ holds family means of the electrophysiology, $\bar{x}_{f_i}$ is the mean expression of cell $i$'s family, and $W$ is a sparse gene-by-feature weight matrix fitted with the reduced-rank, group-sparse penalty used by Scala et al. Because expression is centred within family, $W$ can only explain variation inside families. A gene's gradient score is the norm of its row of $W$, averaged over bootstrap refits. A gene's identity score comes from a sparse multinomial classifier that predicts $f_i$ from $x_i$. Genes that score high on both form a separate overlap group and are reported on their own.
We add a second, classifier-based definition of ambiguity. We train a calibrated cell-type classifier on electrophysiology alone and take the entropy of its predicted class probabilities for each cell. Genes whose within-family expression correlates with that entropy are "boundary genes." This ties the annotation directly to classification ambiguity in the sense used by the cell-typing literature.
The second stage moves the annotations to humans. We map genes to one-to-one human orthologs. As a sanity check that the gradient axis is conserved, we test whether the electrophysiological features it predicts in mouse are among those the shared human–mouse representation of Ophir et al. found transferable. We also test whether the axis orders human Allen Cell Types cells sensibly where paired data permit. Each gene receives a continuous annotation, and we add ±100 kb windows to capture regulatory variants. We then partition SNP-heritability with standard stratified LD-score-style regression, conditioning on a baseline set of functional and constraint annotations. The main quantity is the difference in per-SNP heritability enrichment between gradient and identity annotations, estimated jointly in one model so that the two annotations compete for the same signal.
The third stage takes up the isoform warning. For channel and synaptic genes in the gradient set, we use isoQTL and sQTL results from the developing-brain atlas to ask whether risk loci colocalize with isoform-level regulation more than with total-expression eQTLs. We also refit $W$ at the transcript level wherever the Patch-seq reads (Smart-seq2, which covers full transcripts) support isoform quantification.
Baselines
The naive reference is a heritability model with only the baseline annotations plus an "all genes expressed in cortex" annotation. Any brain-expressed gene set looks enriched for psychiatric heritability, so a gradient annotation that merely matches this reference shows nothing.
The strongest existing competitor is cell-type specificity enrichment. Here genes are ranked by how specifically they are expressed in each transcriptomic type, the top decile forms an annotation, and heritability enrichment is tested per type. This is the default way the field connects single-cell atlases to GWAS, and it is the most likely explanation for any signal we find, because identity and specificity are closely linked. We run it on the same Patch-seq expression matrix, with the same ortholog mapping, window size, baseline model and GWAS summary statistics, and with the same number of genes as our annotations. Beating it means the gradient annotation keeps a significant enrichment when fitted jointly with the specificity annotations, and that adding it raises total explained heritability more than adding another specificity annotation of the same size. We found no reported heritability-enrichment result on Lacuna for Patch-seq-derived gene sets, so these numbers will be produced here rather than quoted.
The strongest result for the gene-to-electrophysiology step is the sparse reduced-rank model of Scala et al. ($R^2 = 0.38$ from 25 genes) and the sparse bottleneck network ($R^2 \approx 0.40$). Our within-family model is a variant of the first. We will report its cross-validated within-family $R^2$ next to these numbers. If it is much lower, the gradient annotation is mostly noise and the downstream test is underpowered.
Three ablations isolate the key ingredient. The first is the decomposition: we use the unsplit sRRR weights, mixing within- and between-family variation, in place of the gradient score. The second is a curated set of all voltage-gated ion channel and synaptic genes, with no Patch-seq weighting. The third is a matched-null gene set drawn to the same distribution of mutational constraint, expression level, specificity and gene length as the gradient set.
Why not just…?
Why not just use cell-type-specific expression gene sets, as everyone already does? This is the real threat. If heritability follows specificity, the gradient annotation adds nothing. The joint model in the Baselines section settles it directly: the gradient set must survive conditioning on specificity. We will also report the case where it does not, because that would mean ambiguity genes are simply specific genes under another name.
Why not just test all ion-channel genes as a pathway? The pathway papers show that curated aggregation works, so this is plausible. However, ion channels are heavily constrained and long, and both properties inflate heritability enrichment on their own. A curated set cannot tell us whether physiology or constraint drives a signal. The constraint-matched null and the curated-channel ablation will show whether Patch-seq weighting adds anything beyond "these are channels."
Why not just use the 25 sRRR genes directly? Twenty-five genes cover too little of the genome for a stable heritability estimate. The sRRR weights also mix family identity with within-family variation, which is exactly the distinction under test. The unsplit-weights ablation measures what the split adds.
Why not just use MRI or other macroscale phenotypes as the intermediate layer? NetWAS shows that this captures mediated risk, but imaging cannot say which cells or which electrical property carry it. Our question concerns the axis of variation inside cell types, which imaging cannot see. Imaging and our approach answer different questions, so we do not claim ours beats it on prediction.
Why not wait for human Patch-seq with genotypes and estimate heritability of cell ambiguity directly? That would be the ideal design. Human patch data are too scarce for it, as the cross-species work notes. Our design reaches the same question indirectly through annotation enrichment, at the cost of relying on mouse-to-human transfer of the gradient axis.
Proposed Experiments and Decisive Checks
The primary experiment builds gradient, identity and boundary annotations from the motor-cortex Patch-seq data, replicates them in the Allen mouse visual cortex data, and tests them jointly against the specificity baseline on summary statistics for schizophrenia, bipolar disorder and autism. As negative controls we use non-brain traits such as serum calcium, a phenotype whose loci each explain little variance, and anthropometric traits. A gradient-specific enrichment that also appears for calcium would point to a constraint artefact.
Three checks decide the outcome. First, gradient genes must agree with known biophysics: their weights should recover the Kv3.1 and Hcn–$I_h$ relationships that the all-active model study confirmed. Second, the gradient-minus-identity enrichment must be positive for psychiatric traits after matching on constraint and specificity, and it must replicate between the motor-cortex and visual-cortex annotations. Third, colocalization in the developing-brain atlas must show whether the risk loci on gradient genes act through isoform or splicing regulation rather than total expression. If they do, the next step is to build isoform-resolved electrophysiology models, which current Patch-seq work does not have.
A null on the second check, with gradient genes no better than matched identity or specificity sets, would be a clean negative result. It would tell cell-typing researchers that within-family ambiguity is not where inherited psychiatric risk concentrates.
References
- Phenotypic variation of transcriptomic cell types in mouse motor cortex
- Sparse Bottleneck Networks for Exploratory Analysis and Visualization of Neural Patch-seq Data
- Classifying Neuronal Cell Types Based on Shared Electrophysiological Information from Humans and Mice
- Single-neuron models linking electrophysiology, morphology and transcriptomics across cortical cell types
- Cross-ancestry, cell-type-informed atlas of gene, isoform, and splicing regulation in the developing human brain
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Pathway analysis of genome-wide data improves warfarin dose prediction
- A pathway-based association analysis model using common and rare variants